PaperlayerView exact revision on arXiv

Source-linked AI summary

Knowledge Distillation for Efficient Acoustic Echo Control

Ernst Seidel, Pejman Mowlaee, Tim Fingscheidt

arXiv:2608.25596v1eess.AS

TL;DR

Machine-learned AEC can outperform classical approaches but is substantially more computationally demanding, and downscaling often reduces performance. This paper uses teacher–student knowledge distillation to recover that loss in compact CGGN16 models, achieving performance comparable to or better than larger ground-truth-trained models at much lower complexity.

  • Problem

    DNN-based AEC requires substantially more computation and memory than classical signal-processing solutions, while reducing model size can compromise performance.

  • Method

    The paper trains significantly downscaled CGGN16 student models using knowledge distillation from a larger teacher and evaluates alternative loss formulations.

  • Results

    The two-step KD approach matches the overall performance of ground-truth-trained CGGN16-M while using 70% fewer parameters and 14% of its computational complexity.

  • Takeaways & Limitations

    KD enables efficient, high-performance AEC with improved near-end speech preservation relative to larger ground-truth-trained models.

Abstract

from arXiv · show

In recent years, many efforts have been made to supersede classical acoustic echo control (AEC) algorithms with more powerful machine-learned approaches. While surpassing the performance of well-established adaptive filters is very much possible, a remaining challenge is computational complexity. Popular architectures, such as convolutional recurrent networks (CRNs), are by multiple orders of magnitude computationally more expensive than classical signal processing solutions. Scaling down such models is usually straight-forward, but it comes at the cost of a notably reduced performance. We show - to the author's knowledge for the first time in AEC - how these performance drops can be successfully alleviated to a large degree by employing an effective knowledge distillation (KD) process, enabling more potent efficient AEC. Our proposed CGGN16 student AEC models show significantly less near-end speech distortion at only 2% of its teacher's computational complexity, surpass the overall performance of a six times more complex model trained on ground-truth labels, and outperform other AEC-focused architectures from recent literature.

I. INTRODUCTION

Machine-learned AEC can improve on classical adaptive filters, but its computational and memory demands make efficient deployment difficult. This work investigates knowledge distillation to recover performance lost when AEC models are downscaled.

  • Motivation: DNN-based AEC requires substantially more computation and memory than classical signal-processing solutions, challenging deployment on edge devices.Modern DNNs may involve millions of parameters and billions of FLOPS.
  • Motivation: Reducing DNN size beyond a certain point typically compromises performance and accuracy, potentially increasing residual echo and near-end speech degradation.Knowledge distillation uses a larger teacher to train a smaller student and has improved performance in related tasks such as noise suppression.
  • Contribution: The paper applies knowledge distillation to a significantly downscaled CGGN16 AEC model to mitigate its performance drop.The study also evaluates different loss formulations and targets performance comparable to larger ground-truth-trained variants.

II. SYSTEM OVERVIEW AND PROPOSED METHOD

The AEC framework simulates a far-end signal traveling through a room and combines its echo with near-end speech and background noise at the microphone. The models process the microphone and far-end signals in the frequency domain.

  • Processing Framework: The far-end reference x(n) propagates through the room and forms the echo d(n) through loudspeaker processing and room convolution.The room impulse response is h(n), while loudspeaker nonlinearities are represented by x′(n) = fNL(x(n)).
  • Processing Framework: The microphone signal is modeled as y(n) = s(n) + n(n) + d(n), combining near-end speech, background noise, and echo.The AEC receives x(n) and y(n) as input signals.
  • Processing Framework: Signals are sampled at 16 kHz, framed into 1024-sample windows with 128-sample shifts, and transformed into frequency-domain representations.An oversampled filterbank with factor 2 is used, with DFT length K = 512.

B. Baseline Architecture

The baseline CGGN16 is a convolutional grouped-GRU encoder–decoder whose size and complexity are controlled mainly by feature-map width and recurrent grouping. A larger F = 64 model serves as the teacher for later distillation.

  • Baseline Architecture: CGGN16 uses three convolutional encoder–decoder blocks that halve spatial feature dimensions while expanding channels to the bottleneck.The architecture is parameterized by the base feature-map count F.
  • Baseline Architecture: At the recurrent bottleneck, feature maps are divided into g groups processed by separate GRUs, reducing each GRU’s input and hidden-unit dimensions.The grouped outputs are combined, decompressed, and passed to the decoder.
  • Baseline Architecture: All CGGN variants retain the same architecture while changing F and g to adjust parameter count and computational complexity.The usual choice g = F/4 provides substantial parameter reduction without significant performance trade-offs.
  • Baseline Architecture: The F = 64 model is used as the teacher, while grouping is generally set to g = F/4.Increasing the base kernel count beyond F = 64 showed no preliminary performance gain.

C. Proposed Knowledge Distillation Losses

The proposed method trains a smaller CGGN16 student with knowledge-distillation losses that use the teacher’s enhanced output, either alone or alongside ground-truth objectives. The study compares time-domain, frequency-domain, and joint formulations.

  • Approach: The work drastically reduces model size and complexity, then uses knowledge distillation to recover performance lost through downscaling.Several distillation approaches are distinguished by their loss functions.
  • Teacher–Student Learning: The teacher’s enhanced output serves as a training target that guides the smaller student network toward improved performance.The teacher and student are distinguished by superscripts T and S, respectively.
  • Loss Formulations: Ground-truth training uses a time-domain logMSE objective, while the student output is evaluated over the entire time sequence and batches.The experiments target echo suppression rather than noise reduction.
  • Loss Formulations: The first distillation approach replaces the ground-truth signal with the teacher output, and a joint loss can combine ground-truth and distillation terms with α = 0.5.The frequency-domain distillation loss provides an alternative to the time-domain formulation.

A. Datasets and Training Details

The study uses disjoint development and test datasets, with the test set varying speakers, noises, nonlinearities, and room acoustics to assess generalization. Models are trained deterministically from scratch under a fixed optimization setup.

  • Datasets: The development set is close to but disjoint from training data, whereas the test set uses different resources and parameterization for unseen-condition generalization.Test speakers come from TIMIT, noise types from the ETSI database, nonlinearities use an arctan function, and room impulse responses come from the Aachen database.
  • Datasets: Signal components are mixed at SER values from −9 to 9 dB and SNR values from 5 to 20 dB.Both ranges use 3 dB increments.
  • Training: All models are causal and trained from scratch on a GTX 1080 Ti GPU using PyTorch2.The training setup uses fixed seeds, Adam optimization, batch size 16, and 200-frame backpropagation-through-time sequences.
  • Training: Training runs are deterministic, using a fixed seed, Adam optimization, batch size 16, and 200-frame backpropagation-through-time unrolling.The passage also specifies training from scratch on a GTX 1080 Ti GPU using PyTorch2.

B. Evaluation Metrics

Evaluation emphasizes challenging double-talk while also covering single-talk conditions, using objective black-box metrics and subjective listening scores. Initial convergence is allowed before metric computation, and evaluated sections exclude added context.

  • Evaluation conditions: Results are mainly discussed for double-talk, while single-talk far-end and near-end conditions are also evaluated.Each condition is assessed with its own subset of metrics.
  • Objective metrics: Black-box metrics isolate enhanced-signal components, including echo, near-end speech, and noise, according to ITU-T Recommendations P.1100 and P.1110.The enhanced signal is decomposed as e(n) = ˜s(n) + ˜d(n) + ˜n(n).
  • Objective metrics: ERLEBB uses a first-order IIR smoothing approximation with α = 0.99 and is averaged over the complete evaluated sequence.Near-end speech preservation is measured with PESQ, PESQBB, LSD, and LPS.
  • Subjective metrics: Subjective evaluation uses crowd-sourced P.808/P.831 listening tests producing DT O* and DT E* MOS scores on noiseless test data.DT O* measures near-end speech quality and DT E* measures echo annoyance.
  • Evaluation procedure: Double-talk sections are preceded by single-talk sections for initial convergence, and added sections are removed before metric calculation.Each added section lasts 8 s to 12 s.

C. Ablations on Knowledge Distillation Strategy

Ablations show that knowledge-distillation losses generally outperform ground-truth-only training for the small CGGN16 student. Frequency-domain KD is strongest among single-step methods, while subsequent ground-truth fine-tuning improves it further.

  • Ablation setup: The small CGGN16 student uses F = 8 base kernels and g = 2 GRU groups in the development-set double-talk ablation.The comparison evaluates loss functions introduced for ground-truth training and knowledge distillation.
  • Single-step KD: Almost all KD losses improve performance over the ground-truth loss JGTt (1).Mixing ground truth with time-domain KD gives a moderate improvement, while adding frequency-domain KD is more potent.
  • Single-step KD: JKDf achieves the best single-step KD performance, ranking first three times while omitting ground-truth labels during training.The result identifies frequency-domain KD as the strongest single-step formulation in this ablation.
  • Two-step KD: The two-step JKDf →JGTt strategy improves performance further and achieves four overall first-place rankings.The second stage fine-tunes the KD-trained model on ground-truth labels.
  • Additional observations: Feature matching failed to improve over the no-KD baseline, while KD models often converged faster and required less training time.The authors suggest the CGGN16 may be too shallow to benefit from intermediate feature alignment.

D. Test Set Results and Discussion

On the test set, KD improves the small CGGN16 across near-end and echo metrics and brings its overall performance close to larger models. Subjective tests similarly favor KD, including comparisons with classical and neural baselines.

  • Objective results: The two-step KD approach improves both near-end and echo-related metrics over CGGN16-S in double-talk and single-talk conditions.Near-end metrics include PESQBB, LPS, ESTOI, and DT O; echo metrics include ERLEBB, DT E, and ST E.
  • Objective results: 70% fewer parameters and 14% of the computational complexity achieve about the same overall performance as ground-truth-trained CGGN16-M.The 14% complexity is described as one-sixth of the larger model’s complexity.
  • Subjective results: Subjective listening tests favor the smaller CGGN16-S over CGGN16-M, especially after the first KD step.KD also reduces distracting residual echo and improves near-end speech quality when echo dominates the microphone signal.
  • Baseline comparisons: Compared with DLAC-Kalman, both KD approaches use less than 4% of its computational requirements while achieving better DT O* and DT E* MOS scores.The comparison concerns subjective double-talk near-end speech quality and echo annoyance.
  • Complexity trade-offs: KD recovers much of the size-related performance loss, with distilled students outperforming the larger CGGN-M in DT O* and LPS.Figure 3 plots DT O* against FLOPS and LPS against parameter count on the test-set double-talk portion.

IV. CONCLUSIONS

Knowledge distillation mitigates performance loss from downscaling AEC models, enabling efficient models with improved near-end speech preservation and better overall performance using 14% of the FLOPS.

  • 14% of the FLOPS enables improved near-end speech preservation and clearly better overall performance than a much larger ground-truth-trained model.These results support efficient, high-performance AEC models through teacher-student learning.
Loading 2608.25596v1…